encoderfile 0.6.2

Distribute and run transformer encoders with a single file.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151

use super::Tensor;
use mlua::prelude::*;
use ndarray::ArrayD;

impl Tensor {
    #[tracing::instrument(skip_all)]
    pub fn clamp(&self, min: Option<f32>, max: Option<f32>) -> Result<Self, LuaError> {
        let input = self
            .0
            .as_slice()
            .ok_or_else(|| LuaError::external("Array must be contiguous"))?;

        let mut out = ArrayD::<f32>::zeros(self.0.raw_dim());
        let out_slice = out
            .as_slice_mut()
            .ok_or_else(|| LuaError::external("Failed to fetch output slice"))?;

        // NaN bound policy: if any bound is NaN, everything becomes NaN. For IEEE-754 compliance :d
        if min.is_some_and(f32::is_nan) || max.is_some_and(f32::is_nan) {
            for dst in out_slice.iter_mut() {
                *dst = f32::NAN;
            }
            return Ok(Self(out));
        }

        match (min, max) {
            (Some(lo), Some(hi)) => {
                for (dst, &src) in out_slice.iter_mut().zip(input.iter()) {
                    *dst = src.max(lo).min(hi);
                }
            }
            (Some(lo), None) => {
                for (dst, &src) in out_slice.iter_mut().zip(input.iter()) {
                    *dst = src.max(lo);
                }
            }
            (None, Some(hi)) => {
                for (dst, &src) in out_slice.iter_mut().zip(input.iter()) {
                    *dst = src.min(hi);
                }
            }
            (None, None) => {
                out_slice.copy_from_slice(input);
            }
        }

        Ok(Self(out))
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_clamp_correctness() {
        let tensor = Tensor(ndarray::array!([-5.0, -1.0, 0.0, 1.0, 5.0]).into_dyn());
        let result = tensor
            .clamp(Some(-1.0), Some(1.0))
            .expect("Failed to clamp");
        let expected = Tensor(ndarray::array!([-1.0, -1.0, 0.0, 1.0, 1.0]).into_dyn());
        assert_eq!(result.0, expected.0);
    }

    #[test]
    fn test_clamp_lower_bound_only() {
        let tensor = Tensor(ndarray::array!([-3.0, 0.0, 2.0]).into_dyn());
        let result = tensor
            .clamp(Some(0.0), None)
            .expect("Failed to clamp tensor");
        let expected = Tensor(ndarray::array!([0.0, 0.0, 2.0]).into_dyn());
        assert_eq!(result.0, expected.0);
    }

    #[test]
    fn test_clamp_upper_bound_only() {
        let tensor = Tensor(ndarray::array!([-3.0, 0.0, 2.0, 5.0]).into_dyn());
        let result = tensor
            .clamp(None, Some(2.0))
            .expect("Failed to clamp tensor");
        let expected = Tensor(ndarray::array!([-3.0, 0.0, 2.0, 2.0]).into_dyn());
        assert_eq!(result.0, expected.0);
    }

    #[test]
    fn test_clamp_infinite_bounds() {
        let tensor = Tensor(ndarray::array!([-3.0, 0.0, 2.0, 5.0]).into_dyn());
        let result = tensor
            .clamp(Some(f32::NEG_INFINITY), Some(f32::INFINITY))
            .expect("Failed to clamp tensor");
        let expected = Tensor(ndarray::array!([-3.0, 0.0, 2.0, 5.0]).into_dyn());
        assert_eq!(result.0, expected.0);
    }

    #[test]
    fn test_clamp_multidimensional() {
        let tensor =
            Tensor(ndarray::array!([[-3.0, 3.0], [0.0, 0.0], [2.0, 2.0], [5.0, 5.0]]).into_dyn());
        let expected_shape = tensor.0.shape().to_owned();

        let result = tensor
            .clamp(Some(0.0), Some(1.0))
            .expect("Failed to clamp tensor");

        let expected =
            Tensor(ndarray::array!([[0.0, 1.0], [0.0, 0.0], [1.0, 1.0], [1.0, 1.0]]).into_dyn());

        assert_eq!(result.0.shape(), expected_shape.as_slice());
        assert_eq!(result.0, expected.0);
    }

    #[test]
    fn test_clamp_identity() {
        let tensor = Tensor(ndarray::array!([-3.0, 0.0, 2.0, 5.0]).into_dyn());
        let result = tensor.clamp(None, None).expect("Failed to clamp tensor");
        assert_eq!(result.0, tensor.0);
    }

    #[test]
    fn test_clamp_min_equals_max() {
        let tensor = Tensor(ndarray::array!([0.0, 3.0, 10.0]).into_dyn());
        let result = tensor
            .clamp(Some(3.0), Some(3.0))
            .expect("Failed to clamp tensor");
        let expected = Tensor(ndarray::array!([3.0, 3.0, 3.0]).into_dyn());
        assert_eq!(result.0, expected.0);
    }

    #[test]
    fn test_clamp_inverted_bounds() {
        let tensor = Tensor(ndarray::array!([0.0, 3.0, 10.0]).into_dyn());
        let result = tensor
            .clamp(Some(5.0), Some(2.0))
            .expect("Failed to clamp tensor");
        let expected = Tensor(ndarray::array!([2.0, 2.0, 2.0]).into_dyn());
        assert_eq!(result.0, expected.0);
    }

    #[test]
    fn test_clamp_nan() {
        // clamping with NaN bounds nuke the entire tensor. Just so that we have no surprises later ;)
        let tensor = Tensor(ndarray::array!([0.0, 3.0, 10.0]).into_dyn());
        let result = tensor
            .clamp(Some(f32::NAN), Some(f32::NAN))
            .expect("Failed to clamp tensor");
        let expected = Tensor(ndarray::array!([f32::NAN, f32::NAN, f32::NAN]).into_dyn());
        for (a, b) in result.0.iter().zip(expected.0.iter()) {
            assert!(a.is_nan() && b.is_nan());
        }
    }
}